Skip to content

Providers

DeltaLLM supports native provider adapters and OpenAI-compatible upstreams. Configure a model deployment with a provider, upstream model ID, credentials, and optional API base. Applications then call the deployment name or route group through DeltaLLM.

Choose an integration path

Upstream type Configuration path
OpenAI, Anthropic, Azure OpenAI, Gemini, Bedrock, or ElevenLabs Use the named provider adapter and its native authentication fields
OpenRouter, Groq, Together AI, Fireworks AI, DeepInfra, or Perplexity Use the named OpenAI-compatible preset
DeepSeek, Z.ai, Qwen, Tencent TokenHub, or MiniMax Use a named compatible chat integration
vLLM, LM Studio, or Ollama Supply the reachable API base for the self-hosted server
Another OpenAI-compatible service Configure its provider/model prefix and explicit API base, then verify the required endpoint and parameters

The generated capability and model catalog is derived from the runtime provider registry. It is the source of truth for the provider modes DeltaLLM currently routes.

Minimal deployment

model_list:
  - model_name: app-chat
    deltallm_params:
      provider: openai
      model: gpt-5-mini
      api_key: os.environ/OPENAI_API_KEY
    model_info:
      mode: chat

Applications request app-chat; the upstream model and credential remain server-side. In production, prefer named credential records or an external secret resolver over embedding provider secrets in a committed configuration file.

Continue to Model deployments for database-managed definitions, credential behavior, pricing metadata, and model modes. Use Routing and failover to combine deployments behind one route group.

Support boundaries

  • Capability is declared per configured model mode, not inferred from marketing names.
  • Provider-specific parameters are forwarded only where the adapter explicitly supports them.
  • Unknown model IDs may work even when they are absent from the curated catalog; test them against the selected adapter.
  • Model catalogs record their verification date and source links. Upstream availability and pricing can change after that date.
  • Provider credentials are write-only operational secrets and should never appear in logs, documentation examples, client code, or public error bodies.